Physics-Embedded Bayesian Neural Network (PE-BNN) to predict Energy Dependence of Fission Product Yields with Fine Structures

This paper introduces a physics-embedded Bayesian neural network (PE-BNN) framework that integrates an energy-independent phenomenological shell factor and WAIC-based hyperparameter optimization to accurately predict energy-dependent fission product yields with fine structures and close agreement with known nuclear shell effects.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Big Picture: Predicting the Unpredictable

Imagine you are a chef trying to predict exactly how a giant, complex cake will crumble when you smash it. In the world of nuclear physics, this "cake" is an atom (like Uranium), and the "smash" is a neutron hitting it, causing it to split (fission).

When an atom splits, it doesn't break into just two equal pieces. It shatters into hundreds of different smaller fragments (called fission products). Scientists need to know exactly how much of each fragment is created, especially when the "smash" happens with different amounts of energy (like a gentle tap vs. a hard hit).

For decades, scientists have had a hard time predicting the tiny, wiggly details of this shattering process, especially when the energy changes. This paper introduces a new "smart chef" that can predict these details with high accuracy.

The Problem: The "Smooth" vs. The "Bumpy"

Think of the data scientists have as a map.

• The Old Map (Linear Interpolation): Traditionally, if scientists knew the result at a "low energy" (like 0.5 MeV) and a "high energy" (like 14 MeV), they just drew a straight line between them to guess what happened in the middle. It's like assuming the road between two cities is perfectly flat.
• The Reality: The road isn't flat. It has hills, valleys, and bumps. These "bumps" are called fine structures. They are caused by the internal "skeleton" of the atom (nuclear shells), which makes certain fragments more likely to form than others.
• The Gap: Existing computer models were good at drawing the general shape of the road (the global trend) but terrible at predicting the specific bumps (the fine structures).

The Solution: The "Physics-Embedded" Smart Chef

The authors built a Physics-Embedded Bayesian Neural Network (PE-BNN). Let's break down that fancy name:

1. Neural Network (The Brain): This is a type of AI that learns by looking at thousands of examples. It's like a student who has read every textbook on nuclear fission.
2. Bayesian (The Confidence Meter): Unlike a standard AI that just gives you one answer, this one gives you an answer and a confidence score. It says, "I think the answer is X, and I'm 95% sure." It's like a weather forecaster who doesn't just say "It will rain," but "There is a 95% chance of rain."
3. Physics-Embedded (The Secret Ingredient): This is the most important part. Usually, you just feed an AI raw data and let it figure everything out. But here, the authors gave the AI a cheat sheet based on real physics.

The "Shell Factor" Analogy

Imagine the atom is a building made of bricks. Some floors are reinforced with steel (these are the nuclear shells). When the building collapses, the reinforced floors hold together differently than the weak ones.

The authors created a special input feature called the "Shell Factor."

• Without the Shell Factor: The AI tries to guess the collapse pattern just by looking at the rubble. It gets the general pile shape right but misses the specific reinforced sections.
• With the Shell Factor: The AI is handed a blueprint that says, "Hey, there's a super-strong steel floor at level 134 and another at 140." Suddenly, the AI can predict exactly where the heavy chunks will land.

How It Works in Practice

The researchers trained this AI on a massive dataset of past experiments and theoretical calculations. They didn't just tell it "predict the yield." They told it:

1. Look at the mass of the fragments.
2. Look at the energy of the incoming neutron.
3. Crucially: Look at the "Shell Factor" (the blueprint of the atom's internal structure).

They also used a special mathematical rule (called WAIC) to act as a strict teacher. This rule prevented the AI from "memorizing" the data (overfitting) and forced it to learn the actual rules of the game.

The Results: Why This Matters

The results were impressive, like a student acing a test they never studied for:

1. Capturing the Bumps: The AI successfully predicted the "fine structures"—the specific peaks and valleys in the data that other models missed.
2. The Energy Trend: It correctly predicted that as the "smash" gets harder (higher energy), the specific "bumps" (shell effects) start to smooth out, just like a rough rock becoming a smooth pebble when tumbled in a river.
3. The "Magic" Connection: Here is the coolest part. The AI was never taught about neutrons (the tiny particles that fly out when the atom splits). It was only taught about the fragments.
   • However, because the AI learned the deep physics of how the atom breaks, it independently figured out that the fragments shift in a way that perfectly matches the known behavior of the flying neutrons.
   • Analogy: Imagine a detective who only looks at the broken glass at a crime scene. Without ever seeing the gun, the detective deduces exactly how hard the bullet hit and what kind of gun was used, just by analyzing the glass patterns. The AI did the same thing: it learned the physics of the split so well that it "discovered" the neutron behavior on its own.

The Takeaway

This paper shows that we don't have to choose between "pure data science" (letting the computer guess) and "pure physics" (using complex equations).

By embedding physics knowledge directly into the AI's brain, we get a tool that is:

• Smarter: It understands the "why" behind the numbers.
• More Accurate: It predicts the tiny details that matter for nuclear reactors and safety.
• Trustworthy: It knows when it's confident and when it's guessing.

This new framework is a bridge between the messy, complex reality of nuclear physics and the clean, powerful predictions of modern AI, helping us design safer reactors and understand the universe better.

Technical Summary

Here is a detailed technical summary of the paper "Physics-Embedded Bayesian Neural Network (PE-BNN) to predict Energy Dependence of Fission Product Yields with Fine Structures."

1. Problem Statement

Fission Product Yields (FPYs) are critical observables for nuclear reactors, astrophysical nucleosynthesis, and nuclear forensics. However, current nuclear data libraries (e.g., JENDL, ENDF) face two major limitations:

• Discrete Energy Points: Libraries typically provide FPY data only at three specific neutron energies (thermal, 0.5 MeV, and 14 MeV). Intermediate energies are estimated via linear interpolation, which recent experiments have shown to be inaccurate due to significant deviations.
• Inability to Reproduce Fine Structures: Neither traditional phenomenological models nor standard machine learning (ML) approaches can simultaneously capture global energy trends and the "fine structures" (local variations in yield due to nuclear shell effects) across different isotopes and incident neutron energies.
• Data Scarcity: Comprehensive experimental investigations are limited by the lack of mono-energetic neutron sources, necessitating robust theoretical models.

2. Methodology

The authors developed a Physics-Embedded Bayesian Neural Network (PE-BNN) framework that integrates prior nuclear physics knowledge directly into the ML architecture.

A. Model Architecture

• Bayesian Neural Network (BNN): The model uses a two-hidden-layer architecture (optimized to 11-11 neurons). It employs the No-U-Turn Sampler (NUTS) within the NumPyro library to approximate the posterior distribution of model parameters $p(W|D)$, allowing for uncertainty quantification (credible intervals).
• Training Data: The dataset comprises 80% of JENDL-5, experimental data from EXFOR, and theoretical model outputs. The remaining 20% of JENDL-5 serves as a validation set. The training targets are post-neutron independent mass yields.

B. Physics-Embedded Input Features

The core innovation is the introduction of a Shell Factor (SF) as an explicit input feature to encode nuclear structure physics:
$$SF = \exp\left(-\frac{E}{kT}\right) \times \sum_{i=1}^{3} W_i (N_i + N_{i+3})$$

• Excitation Energy Damping: The term $\exp(-E/kT)$ represents the Boltzmann factor, where $E$ is excitation energy and $kT \approx 1.5$ MeV (derived from Fermi-gas relations and Langevin simulations). This models the physical reality that shell effects diminish as incident neutron energy increases.
• Shell Structure Encoding: The Gaussian terms $N_i$ are centered at mass numbers $A=134$ (doubly magic), $A=140$, and $A=144$ (deformed shells), with corresponding mirror terms for light fragments. This explicitly guides the network to recognize shell-stabilized regions.
• Hyperparameter Optimization: Model selection and hyperparameter tuning are performed using the Watanabe-Akaike Information Criterion (WAIC) to prevent overfitting and ensure the model generalizes well without relying on iterative parameter retuning for specific reaction systems.

C. Validation Strategy

The model was validated against:

1. Independent experimental cumulative yields (measured via $\gamma$-ray spectroscopy) from recent high-precision studies (e.g., Tonchev et al.).
2. Prompt neutron multiplicity data (which were not used in training) to check for emergent physical consistency.
3. Comparison with the semi-empirical GEF model and linear interpolation of JENDL-5.

3. Key Contributions

• Unified Framework: The PE-BNN successfully bridges the gap between statistical learning and nuclear physics, reproducing both global energy trends and fine structural variations in a single model.
• Explicit Physics Encoding: Unlike "black-box" ML, this approach embeds phenomenological shell information (via the SF) directly into the input layer, making the model's behavior interpretable and physically grounded.
• Emergent Physical Consistency: The model predicts energy-dependent behaviors consistent with prompt neutron multiplicities, despite never being trained on neutron emission data. This demonstrates that the network learned the underlying physical correlations between mass yields and neutron emission from FPY data alone.
• Superior Interpolation: The framework provides accurate predictions for intermediate neutron energies (0.5–14 MeV), outperforming traditional linear interpolation which fails to capture non-linear energy dependencies.

4. Key Results

• Fine Structure Reproduction: Including the Shell Factor (SF) improved the log-likelihood of the model by 21.3% (training) and 34.6% (testing). The model successfully reproduced fine structures in heavy and light peaks for isotopes like $^{235}\text{U}$, $^{232}\text{Th}$, and $^{243}\text{Cm}$.
• Energy Dependence:
  • Heavy Fragments: The model correctly predicted the shift of the heavy peak toward lower mass numbers as incident neutron energy increases. This is attributed to increased prompt neutron emission at higher excitation energies, a behavior consistent with Brosa's fission model.
  • Light Fragments: The model captured the relative stability of light fragment peaks (e.g., $A \approx 99-105$) and their weak energy dependence, consistent with constant prompt neutron multiplicities in this region.
• Shell Effect Hierarchy: The results revealed a hierarchy in shell stabilization: the peak at $A=134$ (doubly magic) remains robust against increasing energy, while peaks at $A=138$ and $A=143$ (deformed shells) dampen and disappear more rapidly.
• Validation: The PE-BNN predictions for $^{235}\text{U}$, $^{238}\text{U}$, and $^{239}\text{Pu}$ fell within the $\pm 1\sigma$ and $\pm 2\sigma$ credible intervals of experimental data across a wide energy range (0.5–14.8 MeV), significantly outperforming linear interpolation and showing better agreement with experimental trends than the GEF model in specific energy regimes.

5. Significance

• Reactor Physics & Safety: The ability to accurately predict FPYs at intermediate energies is crucial for reactor criticality evaluations, burnup calculations, and safety assessments where standard libraries lack data.
• Nuclear Forensics: Improved energy-dependent yield predictions enhance the capability to identify the origin and history of nuclear materials.
• Methodological Advancement: This work demonstrates that "physics-embedded" machine learning is a viable and superior strategy for nuclear data evaluation. It moves beyond simple data fitting to create models that respect fundamental physical laws (shell effects, damping mechanisms) while leveraging the pattern recognition power of Bayesian inference.
• Generalizability: The framework suggests a pathway for applying similar physics-embedded strategies to other complex nuclear observables where systematic features can be encoded as model inputs.